JP2024018002A - Operation management support device and operation management support method - Google Patents

Abstract

[Problem] To support plant operators in speeding up and improving the accuracy of decision-making. An operation management support device 100 inputs measurement data indicating the performance of a plant 210 for predetermined control regarding the operation of a plant 210, and performs predictive control for each control value prediction model from one or more control value prediction models. A control value prediction model calculation unit 150 that calculates a control value prediction model, and a result prediction model calculation unit 160 that inputs a prediction control value for each control value prediction model and calculates a predictive control result for each control value prediction model from the result prediction model. , applied to the plant 210 by comparing the predictive control results of each control value predictive model with a predetermined control target value of the plant 210 and evaluating the accuracy of the control value predictive model that calculated each predictive control result. and an output processing section (display processing section 180) that outputs processing result information including at least predicted control values and evaluation results. [Selection diagram] Figure 1

Description

The present invention relates to an operation management support device and an operation management support method, and is applicable to an operation management support device and an operation management support method that support operation management systems of water and sewage plants, chemical plants, electric power plants, garbage disposal plants, etc. It is suitable.

Water and sewage plants are undergoing consolidation and relocation of facilities in order to respond to issues such as decreasing demand due to population decline, aging facilities, and a lack of human resources. Particularly in the water supply field, the Water Supply Law revised in 2018 states that water utilities are required to make efforts to systematically update water supply facilities from a long-term perspective, and it emphasizes the importance of consolidating, abolishing, and relocating facilities. has been clarified.

Currently, plant management operations such as water and sewage plants, chemical plants, and power plants rely on the visual observation and experience of operators. This is often dealt with by having them stay permanently. However, we need to start thinking about solutions now when a shortage of human resources occurs in the future. If facilities are consolidated or relocated in order to make staffing more efficient, a system will become available for centralized monitoring from an integrated base. In the future, it is expected that an increasing number of business entities will switch to a centralized monitoring system from an integrated base, but as a result of personnel reallocation, the amount of work per operator will increase, and the decision-making ability due to facility management with little experience will increase. There are concerns about a decline in There is also a concern that as the number of skilled operators decreases, it will become more difficult to pass on know-how and techniques to less experienced operators.

In response to the above situation, in order to continue facility maintenance and management operations even after personnel are relocated, the five senses or experience of operators will be replaced by IoT (Internet of Things) and AI (Artificial Intelligence) technologies. Solutions have been provided in recent years. Additionally, systems are also provided that use machine learning techniques such as deep learning to detect anomalies, predict control values and results, and the like.

For example, Patent Document 1 describes a prediction system that aims to more efficiently generate a prediction model with higher accuracy as follows: ``A control calculation unit that executes control calculations for controlling a controlled object; includes a prediction unit that calculates a predicted value by inputting one or more state values that can be referenced into a prediction model, and a prediction model generation unit that generates a prediction model in advance.The prediction model generation unit , means for generating a first prediction model based on a first learning dataset including time series data to be predicted and time series data of one or more corresponding state values; and one or more states. When the actual value of the value satisfies a predetermined condition, the actual value of the one or more state values is added to the first learning dataset, and the second learning dataset is generated. and means for presenting the performance of the first prediction model and the performance of the second prediction model.''

JP2022-28338A

As in the above-mentioned Patent Document 1, systems that use machine learning to predict control values related to plant operation have been proposed. However, while machine learning can create highly accurate control value prediction models within the range of the data used for learning, when the data falls outside the range of the data used for learning, i.e. when conditions have not been proven before. However, there was a possibility that accurate control values could not be predicted. Furthermore, since the contents of the control value prediction model are black boxes, there was a problem with explainability. Furthermore, since each plant had different specifications, there was also the issue that it was difficult to utilize (horizontally deploy) plant knowledge.

For example, Patent Document 1 includes a model that predicts the result of the first control and an outlier detection model that detects when measurement data deviates from past performance. When the outlier detection model detects, a new second result prediction model is created and the performance of the first and second result prediction models is presented. At this time, it is necessary to input the measurement data necessary for the result prediction model, input the control value as a parameter, calculate the control value to output the optimal result, and add processing to present it to the operator. However, because the content of the result prediction model is a black box, problems with explainability remained.

The present invention has been made in consideration of the above points, and aims to propose an operation management support device and an operation management support method that can support speeding up and improvement of accuracy of decision-making by plant operators. It is.

In order to solve this problem, the present invention provides one or more methods for calculating a predicted control value, which is a predicted value of a control value in the control, by inputting measurement data indicating the performance of the plant, regarding a predetermined control regarding the operation of the plant. a control value prediction model, a result prediction model that uses the predicted control value as input to calculate a predictive control result that is a predicted value of the control result of the control, and one or more of the control value prediction models that input the measurement data. a control value prediction model calculation unit that calculates a prediction control value for each control value prediction model from the control value prediction model calculation unit; and a control value prediction model calculation unit that calculates a prediction control value for each control value prediction model calculated by the control value prediction model calculation unit; a result prediction model calculation unit that calculates a predictive control result for each control value prediction model from the result prediction model calculation unit; and a prediction control result for each control value prediction model calculated by the result prediction model calculation unit and a predetermined control target value for the plant. an accuracy evaluation processing unit that determines a suitable control value prediction model to be applied to the plant by evaluating the accuracy of the control value prediction model that calculated each predictive control result; An operation management support device is provided, comprising: an output processing unit that outputs processing result information including at least a predictive control value calculated by a predictive model calculation unit and a result of evaluation by the accuracy evaluation processing unit. Ru.

Further, in order to solve such problems, the present invention provides an operation management support method using an operation management support device that supports an operator's decision making regarding predetermined control regarding the operation of a plant, the operation management support device comprising: one or more control value prediction models that calculate a predicted control value that is a predicted value of the control value in the control using measurement data indicating plant performance as input; and a predicted value of the control result of the control using the predicted control value as input. a result prediction model that calculates a predictive control result, the operation management support device inputs the measurement data and calculates a predicted control value for each control value prediction model from one or more of the control value prediction models. a control value prediction model calculation step for calculating a control value, and the operation management support device inputs the predicted control value for each control value prediction model calculated in the control value prediction model calculation step, and calculates the control value from the result prediction model. A result prediction model calculation step of calculating a predictive control result for each prediction model, and the operation management support device calculates the predictive control result for each control value prediction model calculated in the result prediction model calculation step and the predetermined plant an accuracy evaluation step of determining a suitable control value prediction model to be applied to the plant by comparing the control value prediction model with the control target value and evaluating the accuracy of the control value prediction model that calculated each predictive control result; The operation management support device is characterized by comprising an output step of outputting processing result information including at least the predicted control value calculated in the control value prediction model calculation step and the result of the evaluation by the accuracy evaluation step. An operation management support method is provided.

According to the present invention, it is possible to support plant operators in speeding up and improving the accuracy of decision making.

1 is a block diagram showing a configuration example of an operation management support device 100 according to a first embodiment of the present invention. 3 is a flowchart illustrating an example of a processing procedure of a control value prediction model learning process. 3 is a flowchart illustrating an example of a processing procedure of a control value prediction model accuracy evaluation process. 3 is a flowchart illustrating an example of a processing procedure of a result prediction model accuracy evaluation process. 19 is a diagram showing an example of a processing result screen 190. FIG. It is a block diagram showing an example of composition of operation management support device 300 concerning a 2nd embodiment of the present invention. FIG. 3 is a diagram showing an example of group classification of performance measurement data. 12 is a flowchart illustrating an example of a processing procedure for group-specific control prediction model selection processing.

Embodiments of the present invention will be described in detail below with reference to the drawings.

Note that the following description and drawings are examples for explaining the present invention, and are omitted and simplified as appropriate to clarify the explanation. Furthermore, not all combinations of features described in the embodiments are essential to the solution of the invention. The present invention is not limited to the embodiments, and any application examples that match the idea of the present invention are included within the technical scope of the present invention. Those skilled in the art can make various additions and changes to the present invention within the scope of the present invention. The present invention can also be implemented in various other forms. Unless specifically limited, each component may be plural or singular.

In addition, in the following explanation, processing performed by executing a program may be explained, but the program is executed by at least one or more processors (for example, a CPU) to store predetermined processing as appropriate. Since the processing is performed using resources (for example, memory) and/or interface devices (for example, communication ports), the main body of the processing may be a processor. Similarly, the subject of processing performed by executing a program may be a controller having a processor, a device, a system, a computer, a node, a storage system, a storage device, a server, a management computer, a client, or a host. The main body (for example, a processor) that performs processing by executing a program may include a hardware circuit that performs part or all of the processing. For example, the main body of processing performed by executing a program may include a hardware circuit that performs encryption and decryption, or compression and expansion. A processor operates as a functional unit that implements a predetermined function by operating according to a program. Devices and systems that include processors are devices and systems that include these functional units.

A program may be installed on a device, such as a computer, from a program source. The program source may be, for example, a program distribution server or a computer-readable non-transitory storage medium. When the program source is a program distribution server, the program distribution server includes a processor (for example, a CPU) and a non-temporary storage resource, and the storage resource may further store the distribution program and the program to be distributed. Then, by the processor of the program distribution server executing the distribution program, the processor of the program distribution server may distribute the program to be distributed to other computers. Furthermore, in the following description, two or more programs may be realized as one program, or one program may be realized as two or more programs.

(1) First Embodiment (1-1) Configuration of Operation Management Support Apparatus 100 FIG. 1 is a block diagram showing an example of the configuration of an operation management support apparatus 100 according to the first embodiment of the present invention. The operation management support device 100 shown in FIG. 1 is a device that supports operation management of a plant 210, and includes a track record DB 110, a control value prediction model learning section 120, a result prediction model learning section 130, a control value prediction model selection section 140, It includes a control value prediction model calculation section 150, a result prediction model calculation section 160, an accuracy evaluation processing section 170, and a display processing section 180. Further, the display device 220 is assumed to be a display that can be viewed by an operator (user), and is, for example, a display connected to the operation management support device 100.

The track record DB 110 stores track record measurement data used for learning the control value prediction model and the result prediction model. Actual measurement data is past measurement data obtained from a sensor or the like. The control value prediction model learning unit 120 has a function of learning one or more control value prediction models using performance measurement data stored in the performance DB 110. The control value prediction model is an explanatory model that reflects the knowledge of an operator (more preferably a skilled operator). The result prediction model learning unit 130 has a function of learning a result prediction model using the performance measurement data stored in the performance DB 110. The control value prediction model selection unit 140 has a function of selecting a model with high accuracy of predicted control values (for example, the model with the highest accuracy) based on the learning result of the control value prediction model by the control value prediction model learning unit 120. has. The control value prediction model calculation unit 150 has a function of inputting measurement data to one or more control value prediction models learned by the control value prediction model learning unit 120 and calculating respective predicted control values. The result prediction model calculation unit 160 inputs one or more predictive control values calculated by the control value prediction model calculation unit 150 into the result prediction model learned by the result prediction model learning unit 130, and calculates a predictive control result. Has a function. The accuracy evaluation processing unit 170 evaluates the accuracy of the control value prediction model based on one or more predictive control results calculated by the result prediction model calculation unit 160 and the measured value of the control result measured in the plant 210. Has a function. The display processing unit 180 displays information on the control value prediction model selected by the control value prediction model selection unit 140, the predicted control value calculated by the control value prediction model calculation unit 150, accuracy evaluation by the accuracy evaluation processing unit 170, etc. It has a function of displaying a processing result screen showing on the display device 220. Note that the display processing section 180 is an example of an output processing section that outputs processing result information, and the output form of the output processing section is not limited to display, but may be data output, printing, etc.

Each configuration of the above-mentioned operation management support device 100 will be explained in detail from (1-3) onwards.

The operation management support device 100 is, for example, a computer having a processor such as a CPU (Central Processing Unit), a memory, an auxiliary storage device such as an HDD (Hard Disk Drive), and a network interface. In this computer, the track record DB 110 is realized by, for example, an auxiliary storage device, and includes a control value prediction model learning section 120, a result prediction model learning section 130, a control value prediction model selection section 140, a control value prediction model calculation section 150, and a result prediction model calculation section. Software (programs) that implement the functions of the unit 160, accuracy evaluation processing unit 170, and display processing unit 180 are stored in a memory or an auxiliary storage device. Moreover, in FIG. 1, the communication means between the operation management support device 100, the plant 210, and the display device 220 are not particularly limited. Specifically, for example, it is a wireless LAN (Local Area Network) or a wired LAN, and other communication means may also be used.

(1-2) Plant 210
In the following, a water purification plant will be described as an example of a target plant 210 that can be supported by the operation management support device 100. However, the operation management support device 100 is also applicable to other plants.

A water treatment plant is a facility that purifies river water, dam water, and groundwater, and sends the treated water to factories, households, etc., and is composed of multiple facilities. Specifically, the plurality of facilities include, for example, a landing well, a rapid mixing pond, a floc formation pond, a settling basin, a filtration basin, a disinfection basin, and a water distribution basin, and by performing treatment in each of these facilities, Turbidity is removed from the water to be treated and sterilized.

Specifically, in the landing well, coagulation and sedimentation treatment is performed as a process to remove turbidity. In a typical coagulation-sedimentation treatment at a water purification plant, a chemical called a flocculant, such as polyaluminum chloride or aluminum sulfate, is injected into the water to be treated that flows into the facility. One of the purposes of injecting a drug is to remove insoluble substances suspended in water (hereinafter referred to as turbid substances). Suspended matter is negatively charged in water, but it can be electrically neutralized by adding chemicals. Another purpose of injecting the drug is to facilitate aggregation when neutralized suspended matter collides with each other due to the crosslinking effect between the drugs.

Next, in the rapid mixing pond, chemicals are injected into the water to be treated, and rapid stirring causes the suspended solids to collide with each other to form agglomeration nuclei. Next, in the floc formation pond, the flocs are grown by stirring slowly at a strength that does not destroy the flocs. In the settling tank, the grown flocs are settled and removed.

Here are some examples of factors contributing to the floc aggregation mechanism. As the injection amount of the chemical increases, the crosslinking effect becomes stronger, so the flocs grow faster and the flocculation state becomes better (lower turbidity of treated water). However, if the amount of medicine injected is excessive, the suspended solids will become positively charged, which may result in poor aggregation (increased turbidity of the treated water). As the amount of turbidity in water (turbidity of treated water) increases, the frequency of collisions of suspended solids (or flocculation nuclei, flocs) can increase, so the flocs grow faster. As described above, floc growth is affected by plant operating conditions such as the amount of chemical injection and water quality conditions.

Although the floc aggregation mechanism is common to all water treatment plants, the facility configuration and scale differ depending on the water treatment plant. For example, there are differences in the capacity of settling tanks and the presence or absence of inclined plates. The residence time of the water to be treated in the sedimentation tank is determined depending on the amount of water to be treated and the capacity of the sedimentation tank, but the longer the residence time, the easier it is for small flocs to settle, and the flocs are removed from the water to be treated. Moreover, when there is a slope plate, even if the amount of treated water and the capacity of the settling tank are the same, more flocs are removed from the water to be treated than when there is no slope plate.

Furthermore, in the rapid mixing pond, the flocculant is mixed using a stirring device or the like, but the uniformity of mixing changes depending on the shape of the rapid mixing pond, etc., which affects the subsequent growth of flocs. Furthermore, since the sources of water to be treated differ, the water quality also differs from one water treatment plant to another.

From the above, when a water treatment plant is the target plant 210, the prediction model for the amount of chemical injection (control value) and the turbidity of treated water (control result) is determined by the optimal formula for each water treatment plant (plant 210). There are shapes and coefficients. In addition, even at a single water treatment plant, the prediction model used may change due to long-term changes such as changes in water demand or aging of facilities, or short-term changes such as construction of facilities within the water treatment plant. It may not be optimal. If the operators are experienced, they can respond to changes in the situation within the water treatment plant using their previous knowledge, but if operators with little experience manage the facility (plant 210), It becomes difficult to respond appropriately. Therefore, it is assumed that there may be a case in which maintenance work for the plant 210 cannot be handled using only one prediction model.

Below, each configuration of the operation management support device 100 according to this embodiment will be explained in detail.

(1-3) Actual DB110
The track record DB 110 stores track record measurement data used for learning the control value prediction model and the result prediction model. Specifically, the actual measurement data includes a group of sensors (not shown) that measure water quality items (including at least one of water quality, water volume, and operation amount) related to water treatment installed in the plant 210 and operating parameters of equipment. etc. Sensors installed in the plant 210 of the water treatment plant include a turbidity meter that measures water turbidity, a water temperature meter that measures water temperature, a pH meter that measures pH, an alkalinity meter that measures alkalinity, and There are TOC meters and ultraviolet absorbance meters that measure organic matter. Additionally, a water meter that measures the amount of water processed, but not the quality of the water, also falls under the above-mentioned sensor. In this embodiment, the means for measuring water quality items includes at least a turbidity meter (for treated water and treated water), and may also include a sensor capable of measuring water quality items that affect water treatment. Furthermore, other than the above-mentioned sensors, the sensor is not particularly limited as long as it is involved in coagulation or precipitation processing. Note that the performance measurement data is stored in the performance DB 110 as time series information.

(1-4) Control value prediction model learning unit 120
The control value prediction model learning unit 120 uses the performance measurement data stored in the performance DB 110 to execute a "control value prediction model learning process" for learning one or more control value prediction models.

The minimum number of control value prediction models required by the operation management support device 100 according to the present embodiment is one given from the beginning of the operation of the water purification plant. In that case, other models are added when a given model is no longer a suitable model.

The control value prediction model may be an empirical formula, a statistical model, or a physical model that reflects the operator's knowledge, and any model can be used as long as the content can be recognized by the operator (has explanatory properties). good.

ここで、ｍ１、ｍ２、及びｍ３は係数であり、ＣＲ１は式１から得られた処理水量あたりの凝集剤の注入量を示す凝集剤注入率であり、Ｔｕ０は被処理水の濁度である。 For example, as a control value prediction model used in a general water purification plant, there is a control value prediction model M1 shown in Equation 1 below.

Here, m1, m2, and m3 are coefficients, CR1 is the flocculant injection rate indicating the amount of flocculant injected per amount of treated water obtained from Equation 1, and Tu0 is the turbidity of the water to be treated. .

In Equation 1, as the water turbidity Tu0 increases, the flocculant injection rate CR1 increases, but the coefficient m2 takes a value between 0 and 1, so it is not linear. This is because, as described above, as the turbidity of the water to be treated increases, the frequency of collisions of suspended solids (or flocculation nuclei, flocs) increases, and flocs grow even if the crosslinking effect of the flocculant is small.

ここで、ｎ１、ｎ２、ｎ３、及びｎ４は係数であり、ＡＬは被処理水のアルカリ度であり、ＣＲ２は式２から得られた処理水量あたりの凝集剤の注入量を示す凝集剤注入率であり、Ｔは水温である。 Further, for example, as a control value prediction model that considers water quality such as water temperature and alkalinity other than the turbidity of the water to be treated, there is a control value prediction model M2 shown in the following equation 2.

Here, n1, n2, n3, and n4 are coefficients, AL is the alkalinity of the water to be treated, and CR2 is the flocculant injection rate indicating the amount of flocculant injected per amount of treated water obtained from equation 2. , and T is the water temperature.

As explained in Equation 1, in Equation 2, the turbidity Tu0 of the water to be treated has a positive correlation with the flocculant injection rate CR2. On the other hand, water temperature T and alkalinity AL have a negative correlation with flocculant injection rate CR2. This is because the water temperature affects the reaction rate at which the flocculant reacts with alkaline components and changes into positively charged aluminum hydroxide, and lowering the water temperature has a negative effect on the flocculation effect. Furthermore, if alkalinity is insufficient or excessive, it will have a negative effect on the coagulation effect. When the minimum level of alkalinity (10 mg-CaCO3/L) is secured, a negative correlation due to excess alkalinity is considered to be observed. Although not considered in this explanation, it is known that pH also affects the aggregation effect. A flocculant has an effective pH range; for example, in the case of sulfuric acid, which is an example of a flocculant, the effective pH range is 5.8 to 7.8. If the pH is out of the effective range, it will have a negative impact on the aggregation effect.

In the following, a case will be described as a preferred embodiment in which two control value prediction models (for example, M1 in Equation 1 and M2 in Equation 2) exist. The shape of the equation of the control value prediction model is not limited to Equation 1 and Equation 2 above, as long as it has explainability and can predict the flocculant injection rate. Furthermore, the operation management support device 100 is applicable even when three or more control value prediction models exist.

(1-4-1) Control value prediction model learning process FIG. 2 is a flowchart illustrating an example of the processing procedure of the control value prediction model learning process. The control value prediction model learning process is executed by the control value prediction model learning section 120.

According to FIG. 2, first, the control value prediction model learning unit 120 calls one control value prediction model stored in a recording medium (not shown) (step S101).

Next, the control value prediction model learning unit 120 acquires performance measurement data used in the control value prediction model called in step S101 from the performance DB 110 (step S102). Specifically, for example, when the control value prediction model M1 of Equation 1 is called in step S101, the turbidity Tu0 of the water to be treated and the flocculant injection rate CR1 are acquired as actual measurement data, and the control value prediction model M1 of Equation 2 is called in step S101. When the control value prediction model M2 is called, the turbidity Tu0 of the water to be treated, the alkalinity AL of the water to be treated, the water temperature T, and the flocculant injection rate CR2 are acquired as performance measurement data. The period of data acquired in step S102 is not particularly limited, and may be, for example, all the periods stored in the track record DB 110, or may be a predetermined period set in advance.

Next, the control value prediction model learning unit 120 uses the performance measurement data acquired in step S102 to learn the control value prediction model called in step S101, and acquires coefficients (step S103). The learning method of the control value prediction model differs depending on the control value prediction model and is set in advance.

制御値予測モデル学習部１２０は、式３においてＬｏｇ（Ｔｕ０）とＬｏｇ（ＣＲ１－ｍ３）を線形回帰することで、傾きｍ２と切片Ｌｏｇ（ｍ１）を算出することができる。 For example, in the case of the control value prediction model M1 shown in Equation 1, the coefficient m3 is the value of the flocculant injection rate (coagulation (lower limit value of the drug injection rate), and a setting value stored in advance in a recording medium (not shown) is acquired. Equation 1 can be transformed into Equation 3 below.

The control value prediction model learning unit 120 can calculate the slope m2 and the intercept Log(m1) by performing linear regression on Log(Tu0) and Log(CR1-m3) in Equation 3.

For example, in the case of the control value prediction model M2 shown in Equation 2, the objective variable is the flocculant injection rate CR2, the explanatory variables are the turbidity of the water to be treated Tu0, the alkalinity AL of the water to be treated, and the water temperature T. By doing so, the slopes n1, n2, n3 and intercept n4 of each explanatory variable can be calculated.

In addition, in the linear regression or multiple regression described above, the control value prediction model learning unit 120 does not use actual measurement data as is, but uses values normalized using the maximum and minimum values of each explanatory variable. These pretreatment methods are not particularly limited. Furthermore, when performing normalization and standardization, it is necessary to perform similar preprocessing when inputting measurement data (current time information) of the plant 210 and calculating respective predictive control values.

Next, the control value prediction model learning unit 120 stores the coefficients calculated in the learning at step S103 in a recording medium (not shown) (step S104).

Then, the control value prediction model learning unit 120 determines whether there is an unlearned control value prediction model (step S105). If there is an unlearned control value prediction model (YES in step S105), the control value prediction model learning unit 120 repeats the processes of steps S101 to S104 for the control value prediction model. If all the control value prediction models have been learned (NO in step S105), the control value prediction model learning unit 120 ends the control value prediction model learning process.

(1-5) Result prediction model learning unit 130
The result prediction model learning unit 130 learns a result prediction model using the performance measurement data stored in the performance DB 110. In this embodiment, the result prediction model is created using machine learning (artificial neural network, clustering, etc.) that does not reflect the operator's knowledge but can predict control results with high accuracy within the actual performance range. However, as long as the control results can be predicted with high accuracy, the result prediction model may be a statistical model, a physical model, or the like that is a method other than machine learning. Below, as an example, a case will be described in which a result prediction model is learned using an artificial neural network.

The artificial neural network is a mathematical model in which "processing units that linearly transform inputs" are combined on the network, and the result prediction model learning unit 130 performs regression prediction of treated water turbidity based on actual measurement data. At this time, the input performance measurement data is an explanatory variable related to the turbidity of the treated water when the objective variable is the turbidity of the treated water, and specifically, the turbidity of the treated water, These include alkalinity, pH, water temperature, and flocculant injection rate. Note that the explanatory variable is not limited to the above example as long as it can predict the turbidity of the treated water.

Generally, an artificial neural network consists of an input layer that sends explanatory variables to the network, an intermediate layer that receives the explanatory variables from the input layer and performs calculations, and an output layer that outputs the calculation results from the network. Deep learning is an artificial neural network with three or more hidden layers, and the middle layer is a fully connected layer that calculates using all explanatory variables, and a fully connected layer that connects the next layer to prevent overfitting. It consists of a dropout layer with randomly cut parts.

In this embodiment, the result prediction model learning unit 130 first outputs an inferred value of treated water turbidity from the last fully connected layer of the artificial neural network using actual measurement data. Next, the result prediction model learning unit 130 updates the coefficients of the connection weights in order from the last fully connected layer to the first fully connected layer based on the error between the inference value and the correct data (actual measured data of treated water turbidity). (error backpropagation). The result prediction model learning unit 130 then outputs the inferred value of the treated water turbidity again and learns the result prediction model by repeating the process until the error becomes sufficiently small.

(1-6) Control value prediction model selection unit 140
The control value prediction model selection unit 140 selects a control value prediction model with high accuracy of predicted control values from the results of learning one or more (in this example, two) control value prediction models by the control value prediction model learning unit 120. do. The control value prediction model selection unit 140 uses the selected control value prediction model to predict the control value of a predetermined control (specifically, for example, the control value of the operation of the flocculant injection pump) in the operation of the plant 210. It may be applied as a control value prediction model to be used.

Here, the accuracy of the predicted control value refers to the average absolute error (МAE :Mean Absolute Error). It is preferable that the performance measurement data of the flocculant injection rate used here includes performance measurement data after model learning. If performance measurement data after this model learning has been accumulated in the performance DB 110, additional performance measurement data may be acquired from the performance DB 110 when calculating the average absolute error; if it is not accumulated in the performance DB 110, It may be acquired from the plant 210 when calculating the average absolute error.

ここで、Ｎは学習に用いたデータ数であり、ＰＣＲは制御値予測モデルより算出された凝集剤注入率の予測制御値である。平均絶対誤差の算出は、式４に限定されるものではなく、他にも例えば、学習に用いた実績計測データの凝集剤注入率の最大値及び最小値を用いて正規化した値を用いるようにしてもよい。また、精度の指標には、平均絶対誤差率（ＭＡＥＲ：Mean Absolute Error Rate）等を用いてもよく、予測制御値の精度の指標となるものであれば特に限定されない。以下では、制御値予測モデル選択部１４０が「予測制御値の精度が高い制御値予測モデル」として制御値予測モデルＭ２を選択したとして、説明を続ける。 The control value prediction model selection unit 140 compares the average absolute errors calculated from the control value prediction models M1 and M2 described above, and chooses a control value prediction model with a small average absolute error as a control value prediction with a high precision of the control value. Select as model. The mean absolute error (MAE) is calculated using Equation 4 below, for example.

Here, N is the number of data used for learning, and PCR is a predicted control value of the flocculant injection rate calculated from the control value prediction model. Calculation of the average absolute error is not limited to Equation 4. For example, it may be possible to use a value normalized using the maximum and minimum values of the flocculant injection rate of the actual measurement data used for learning. You can also do this. Moreover, the mean absolute error rate (MAER) or the like may be used as the accuracy index, and is not particularly limited as long as it is an index of the accuracy of the predictive control value. The following explanation will be continued assuming that the control value prediction model selection unit 140 selects the control value prediction model M2 as the "control value prediction model with high accuracy of predicted control values."

(1-7) Control value prediction model calculation unit 150
The control value prediction model calculation unit 150 adds the measured data of the plant 210 (current time information) and calculate each predictive control value.

Further, the control value prediction model calculation unit 150 calculates the predicted control value calculated from the control value prediction model (control value prediction model M2 in this example) selected as a highly accurate control value prediction model by the control value prediction model selection unit 140. is transmitted to the monitoring and control system of the plant 210. The supervisory control system of the plant 210 then controls the operation of the flocculant injection pump of the plant 210 using the received predicted control value. Note that the supervisory control system of the plant 210 may not operate the received predicted control values as they are, but may operate them after confirmation by an operator.

(1-8) Result prediction model calculation unit 160
The result prediction model calculation unit 160 calculates the predicted control values calculated by the control value prediction model calculation unit 150 (in this example, the predicted control values of two flocculant injection rates CR1 and CR2), the turbidity of the water to be treated, and the turbidity of the water to be treated. Explanatory variables such as water alkalinity, pH, and water temperature are input to the result prediction model learned by the result prediction model learning unit 130, and predictive control results for each treated water turbidity are calculated.

(1-9) Accuracy evaluation processing unit 170
The accuracy evaluation processing unit 170 uses the predictive control results (in this example, two predicted control values of treated water turbidity) calculated by the result prediction model calculation unit 160 and the control target of treated water turbidity measured in the plant 210. A "control value prediction model accuracy evaluation process" is executed to evaluate the accuracy of the control value prediction model at the current time based on the value. Note that the accuracy here refers to the absolute error or absolute error rate between the predictive control result of treated water turbidity and the control target value, and the control target value is set in advance for each plant 210 (water treatment plant) and is It is assumed that the data is stored in the illustrated recording medium.

(1-9-1) Control Value Prediction Model Accuracy Evaluation Process FIG. 3 is a flowchart showing an example of the processing procedure of the control value prediction model accuracy evaluation process. The control value prediction model accuracy evaluation process is executed by the accuracy evaluation processing section 170.

According to FIG. 3, first, the accuracy evaluation processing unit 170 calls a result prediction model stored in a recording medium (not shown) (step S201).

Next, the accuracy evaluation processing unit 170 acquires measurement data that is measured in the plant 210 and is used in the result prediction model (step S202). Specifically, the measurement data to be acquired is, for example, the turbidity of the water to be treated, the alkalinity of the water to be treated, the pH, and the water temperature.

Next, the accuracy evaluation processing unit 170 acquires the predicted control value from the control value prediction model calculation unit 150 (step S203). In step S203, the accuracy evaluation processing unit 170 acquires predicted control values for each control value prediction model (M1, M2). In this example, it is assumed that the predicted control value (first predicted control value) of the control value prediction model M1 is first obtained.

Next, the accuracy evaluation processing unit 170 inputs the measurement data acquired in step S202 and the predictive control value (first predictive control value) acquired in step S203 to the result prediction model called in step S201, and makes a prediction. A control result (first predictive control result) is calculated (step S204).

Next, the accuracy evaluation processing unit 170 calculates the absolute error (first absolute error) between the control target value stored in the recording medium and the predictive control result (first predictive control result) calculated in step S204. Calculate (step S205).

Next, the accuracy evaluation processing unit 170 compares the absolute error (first absolute error) calculated in step S205 with a preset threshold (first threshold), and determines whether the absolute error is larger than the threshold. (Step S206).

If the absolute error is larger than the threshold in step S206 (YES in step S206), the predictive control result deviates from the control target value, so the predictive control value input in the calculation of the predictive control result is controlled. It is possible that the learning of the value prediction model (that is, the control value prediction model M1 under evaluation) is insufficient. In this case, the accuracy evaluation processing unit 170 instructs the display processing unit 180 to present a recommendation for relearning the control value prediction model (control value prediction model M1) under evaluation on the display device 220 (step S207). After the processing in step S207, the process advances to step S208.

On the other hand, if the absolute error is less than or equal to the threshold in step S206 (NO in step S206), since the predictive control result is sufficiently close to the control target value, the predictive control value input in calculating the predictive control result is calculated. The learning of the control value prediction model is considered to be sufficient. In this case, since relearning of the control value prediction model (control value prediction model M1) under evaluation is not necessary, the accuracy evaluation processing unit 170 skips step S207 and proceeds to step S208.

In step S208, the accuracy evaluation processing unit 170 determines whether there is a control value prediction model for which the predicted control value has not been acquired in step S203. In this example, the predicted control value (first predicted control value) of the control prediction model M1 has been acquired, but the predicted control value (second predicted control value) of the control prediction model M2 has not yet been acquired. In this case, the accuracy evaluation processing unit 170 determines that there is an unobtained control value prediction model (YES in step S208), returns to step S201 and repeats the process, and in step S203 after the repetition, the second predictive control Get the value. If the predicted control values of all the control prediction models have been acquired, it is determined that there is no control value prediction model that has not yet been acquired (NO in step S208), and the process proceeds to step S209.

In step S209, the accuracy evaluation processing unit 170 compares the predicted control value calculated in step S205 with respect to the currently selected control value prediction model (the highly accurate control value prediction model selected by the control value prediction model selection unit 140). The absolute error with the control target value is compared with the absolute error between the predicted control value and the control target value calculated in step S205 for other control value prediction models. In this example, the "currently selected control value prediction model" is the control value prediction model M2, and the "other control value prediction model" is the control value prediction model M1.

Therefore, in step S209, the accuracy evaluation processing unit 170 calculates the absolute error (second absolute error) between the predicted control value (second predicted control value) in the control value prediction model M2 and the control target value using the control value prediction model M2. Compare with the first absolute error in M1.

As a result of the comparison in step S209, if the absolute error (first absolute error) in another control value prediction model is smaller than the absolute error (second absolute error) in the currently selected control value prediction model (step (YES in S209), the accuracy evaluation processing unit 170 determines that another control value prediction model (control value prediction model M1) has higher accuracy than the currently selected control value prediction model (control value prediction model M2). to decide. Therefore, in this case, the accuracy evaluation processing unit 170 instructs the display processing unit 180 to present a recommendation for changing from the currently selected control value prediction model to the other control value prediction model on the display device 220 ( Step S210). After the process of step S210, the control value prediction model accuracy evaluation process ends.

On the other hand, as a result of the comparison in step S209, if the absolute error (first absolute error) in another control value prediction model is not smaller than the absolute error (second absolute error) in the currently selected control value prediction model In other words, if the absolute error (second absolute error) in the currently selected control value prediction model is less than or equal to the absolute error (first absolute error) in another control value prediction model (NO in step S209) , the accuracy evaluation processing unit 170 determines that the accuracy of the other control value prediction model (control value prediction model M1) is not higher than the currently selected control value prediction model (control value prediction model M2). . Therefore, in this case, the accuracy evaluation processing unit 170 instructs the display processing unit 180 to present the status quo of the currently selected control value prediction model on the display device 220 (step S211). After the process of step S211, the control value prediction model accuracy evaluation process ends.

As described above, the accuracy evaluation processing unit 170 can evaluate the accuracy of the control value prediction model at the current time. Note that in the processing procedure example of FIG. 3, the absolute error calculated in step S205 is used as an evaluation index, but the evaluation index in this embodiment is not limited to this. For example, instead of considering only one point at the current time as shown in Figure 2, calculate the average absolute error or average absolute error rate between the predicted control result of treated water turbidity and the control target value over a certain period of time. However, this may be used as an index for comparative evaluation.

(1-9-2) Result prediction model accuracy evaluation process In addition to the control value prediction model accuracy evaluation process described with reference to FIG. "Model accuracy evaluation processing" is also executed. After calculating the predictive control result in step S204 of the control value prediction model accuracy evaluation process shown in FIG. 3, the accuracy evaluation processing unit 170 executes the result prediction model accuracy evaluation process. The control value prediction model accuracy evaluation process and the result prediction model accuracy evaluation process may be executed separately.

FIG. 4 is a flowchart illustrating an example of the processing procedure of the result prediction model accuracy evaluation process.

According to FIG. 4, first, the accuracy evaluation processing unit 170 determines that the predictive control result calculated in step S204 of FIG. It is determined whether the result is the predictive control result (step S301). In this example, the accuracy evaluation processing unit 170 inputs the measurement data of the plant 210 and the predicted control value (second predicted control value) of the control value prediction model M2 into the result prediction model in step S204 of FIG. Assume that the predictive control result (second predictive control result) of the control value predictive model M2 has been calculated. In this case, the control value prediction model M2 used to calculate the second predictive control result is the control value prediction model currently used in the plant 210 (YES in step S301), and the process proceeds to step S302 to predict the result. Start evaluating the accuracy of the model. On the other hand, if the predictive control result calculated in step S204 in FIG. 3 is not the predictive control result of the control value prediction model currently used in the plant 210 (NO in step S301), the process returns to step S201 in FIG. When the predictive control result is calculated, the process of step S301 is performed again.

In step S302, the accuracy evaluation processing unit 170 determines the absolute error ( In this example, the third absolute error) is calculated.

Next, the accuracy evaluation processing unit 170 compares the absolute error calculated in step S302 (third absolute error) with a preset threshold (second threshold), and determines whether the absolute error is larger than the threshold. (Step S303).

If the absolute error is larger than the threshold in step S303 (YES in step S303), the predictive control result calculated by the result prediction model using the predictive control value of the currently selected control value prediction model M2 is calculated from the actual measurement data. Since there is a discrepancy, it is possible that the learning of the result prediction model that calculated this predictive control result (that is, the result prediction model called in step S201 in FIG. 3) is insufficient. In this case, the accuracy evaluation processing unit 170 instructs the display processing unit 180 to present a recommendation for relearning the result prediction model on the display device 220 (step S304). When a recommendation for relearning the result prediction model is presented on the display device 220, the operator performs a predetermined operation to cause the result prediction model learning unit 130 to relearn the result prediction model. Note that the operation management support device 100 may automatically instruct relearning of the result prediction model through program processing. After the process in step S304, the accuracy evaluation processing unit 170 ends the result prediction model accuracy evaluation process.

On the other hand, if the absolute error is less than or equal to the threshold in step S303 (NO in step S303), the predictive control result calculated by the result prediction model using the predictive control value of the currently selected control value prediction model M2 is actually Since it is sufficiently close to the measured data of , it is considered that the learning of the prediction model based on the calculation of this predictive control result is sufficient. In this case, the accuracy evaluation processing unit 170 instructs the display processing unit 180 to present the status quo of the result prediction model on the display device 220 (step S305), and ends the result prediction model accuracy evaluation process.

(1-10) Display processing section 180
The display processing unit 180 generates a processing result screen 190 that shows processing results (for example, calculated predictive control values, accuracy evaluation, etc.) by each processing unit of the operation management support device 100, and displays and outputs this on the display device 220. do.

FIG. 5 is a diagram showing an example of the processing result screen 190.

In the processing result screen 190 shown in FIG. 5, in an area 191, the contents to be presented to the operator determined by the accuracy evaluation processing section 170 are displayed in order from the latest time. For example, in area 191 of FIG. 5, a recommendation is made to change from control value prediction model M2 to control value prediction model M1, corresponding to the process of step S210 of FIG.

In area 192, a trend graph of the predicted control value of the flocculant injection rate and a trend graph of the predicted control result of the treated water turbidity are displayed. In the area 192, data of all control value prediction models used may be displayed, or any control value prediction model may be selected and displayed (selected by the user or automatically selected). good. However, it is preferable to display at least the measurement data acquired from the plant 210 and the currently selected main control value prediction model (eg, control value prediction model M2). Further, the time width of the trend can be arbitrarily selected or set.

In area 193, items whose trends are displayed in area 192 are displayed in a list. For example, according to the area 193 in FIG. 5, it can be seen that the measurement data, the control value prediction model M1, and the control value prediction model M2 are displayed in the area 192 as a trend graph.

In the area 194, preset values (eg, control target value, first threshold, and second threshold) used in the accuracy evaluation by the accuracy evaluation processing unit 170 are displayed. Note that the processing result screen 190 may be configured so that the operator can change the set value to another value by operating the area 194.

Further, in the area 194, the states of the control value prediction model and the result prediction model are also displayed. To explain using specific values, in the case of FIG. 5, referring to the trend graph in area 192, at the current time ("00:55" according to area 191), the predictive control result (first (predictive control result) is "1.18 degrees," the predictive control result by the control value prediction model M2 (second predictive control result) is "1.29 degrees," and the measured data is "1.22 degrees." As shown in area 194, since the control target value is set to "1.0 degrees," the first absolute error is "0.18 degrees," and the second absolute error is "0.29 degrees." ”. In other words, the absolute error of the control value prediction model M1 is smaller than that of the currently selected main control value prediction model M2. Recommendations are presented. Note that since each absolute error is within "0.3 degrees" of the first threshold value, relearning of the control value prediction model is not recommended. Further, the absolute error (third absolute error) between the predictive control result (second predictive control result) of the control value prediction model M2 used in the plant 210 and the measured data is "0.07 degree", Since it is within "0.2 degrees" of the second threshold, it is not recommended to retrain the result prediction model. Based on the above, it is displayed in the area 194 that the control value prediction model is in a state where "change" is recommended, and the result prediction model is in a state where "maintaining the status quo" is sufficient. Note that the method for displaying the status of the control value prediction model and the result prediction model in area 194 may be any form that can notify the operator of the respective status, such as "change" or "maintain status quo" as shown in FIG. It is not limited to display. For example, the absolute error or error calculated in the explanation using the above specific example may be displayed.

As described above, according to the operation management support device 100 according to the first embodiment, the predetermined control necessary for operating the facility (plant 210) can be explained with the knowledge of the operator reflected. Using a certain control value prediction model, the control value prediction model calculation unit 150 calculates a predicted control value, and uses machine learning to apply the predicted control value to a result prediction model that can predict highly accurate control results within the actual range. is input, the result prediction model calculation unit 160 calculates a predictive control result, and the accuracy evaluation processing unit 170 evaluates the control value prediction model based on the difference between the predicted control value and the control target value. It is possible to assist the operator in easily selecting a highly accurate control value prediction model that can calculate predicted control values with high reliability. As a result, it is possible to support the speed and accuracy of decision-making by operators with little experience in facility management, even if the operators are not experts.

More specifically, if the control value is the amount of chemical used (for example, flocculant injection rate) to be injected into the plant 210, the operation management support device 100 increases the predicted control value of the appropriate flocculant injection rate. By being able to present operators with a control value prediction model that can be calculated with high accuracy, it is possible to early determine and prevent insufficient or excessive amounts of chemicals being used, and support optimization of chemical usage. . Furthermore, if the amount of chemicals used can be optimized, the frequency of replenishing and transporting chemicals will be reduced, which can also be expected to reduce the amount of carbon dioxide gas emitted during manufacturing and transportation of chemicals.

Further, according to the operation management support device 100 according to the first embodiment, the accuracy evaluation processing unit 170 calculates the results based on the difference between the predictive control results calculated by the result prediction model calculation unit 160 and the actual measurement data in the facility. By evaluating the prediction model, it is determined whether the accuracy of the currently selected result prediction model is sufficient, and if the accuracy is insufficient, the result prediction model learning unit 130 retrains the result prediction model. Therefore, the accuracy of the result prediction model for calculating the predictive control result can be maintained at a high level.

Thus, according to the operation management support device 100 according to the present embodiment, a digital twin that can predict highly accurate results within the actual performance range is constructed, and the operator's knowledge is reflected in the prediction of control values to improve explainability. By using a certain model, even a small number of operators can support the efficient maintenance and management of multiple plants.

(2) Second Embodiment FIG. 6 is a block diagram showing a configuration example of an operation management support device 300 according to a second embodiment of the present invention. The operation management support device 300 shown in FIG. 6 has a configuration in which a performance data classification unit 310 is added to the operation management support device 100 according to the first embodiment shown in FIG. Description of the configuration and its processing will be omitted.

The performance data classification unit 310 groups predetermined data based on the similarity between the performance measurement data (turbidity of water to be treated, alkalinity of water to be treated, water temperature, pH, etc.) stored in the performance DB 110. Classify into. The grouping method is, for example, machine learning using the k-means method, but other machine learning, statistical processing, or the like may be used, and is not particularly limited.

FIG. 7 is a diagram showing an example of group classification of performance measurement data. The classification graph 320 in FIG. 7 graphically displays an example of the results of classifying a plurality of performance measurement data stored in the performance DB 110 into groups based on the turbidity and water temperature of the water to be treated. In the classification graph 320, the horizontal axis is the turbidity of the water to be treated, and the vertical axis is the water temperature.

In the case of the classification graph 320 in FIG. 7, Group 1 has both low turbidity and water temperature of the treated water, Group 2 has high turbidity of the treated water and low water temperature, Group 3 has low turbidity of the treated water and high water temperature, Actual measurement data is classified into four groups, including group 4, in which both the turbidity and temperature of the water to be treated are high.

Furthermore, the performance data classification unit 310 can also classify predicted control values calculated from the control value prediction model into groups using the same group classification as described above. In the second embodiment, a group-based control prediction model selection process is performed in which the control value prediction model selection unit 140 selects a control value prediction model with high accuracy of predicted control values for each group classified by the performance data classification unit 310. Execute.

FIG. 8 is a flowchart illustrating an example of a processing procedure for group-specific control prediction model selection processing.

According to FIG. 8, first, the control value prediction model selection unit 140 calls one control value prediction model stored in a recording medium (not shown) (step S401). In this example, two control value prediction models M1 and M2 are stored, and the following explanation assumes that the control value prediction model M1 is called first, but the number of control value prediction models stored in the recording medium is The number of control value prediction models may be three or more, and the order of calling the control value prediction models may be arbitrary.

Next, the control value prediction model selection unit 140 acquires performance measurement data used in the control value prediction model called in step S401 from the performance DB 110 (step S402). The process of step S402 is similar to the process of the control value prediction model learning unit 120 described in step S102 of FIG.

Next, for the control value prediction model (control value prediction model M1) called in step S401, the control value prediction model selection unit 140 acquires the coefficients of the control value prediction model stored in a recording medium (not shown). (Step S403). The coefficient acquired in step S403 is the coefficient stored in the recording medium (not shown) in step S104 of FIG.

Next, the control value prediction model selection unit 140 inputs the performance measurement data acquired in step S402 and the coefficient acquired in step S403 to the control value prediction model (control value prediction model M1) called in step S401, A predictive control value (first predictive control value) is calculated (step S404). As in the first embodiment, when the plant 210 is a water purification plant, the predicted control value calculated in step S404 is, for example, a predicted control value of the flocculant injection rate.

Next, the control value prediction model selection unit 140 classifies the predicted control value (first predicted control value) calculated in step S404 according to the group classification by the performance data classification unit 310 (step S405).

Next, the control value prediction model selection unit 140 calculates the mean absolute error (MAE) between the predicted control value calculated in step S404 and the performance measurement data acquired in step S402 for each group determined by the performance data classification unit 310. (Step S406). Here, the formula for calculating the average absolute error may be, for example, the formula 4 described above, but similarly to the explanation in the first embodiment, the maximum value and minimum value of the flocculant injection rate of the actual measurement data used for learning It is also possible to use a value normalized using a value. Moreover, the average absolute error rate (MAER) or the like may be used instead of the average absolute error as long as it serves as an index of the accuracy of the predictive control value.

Next, the control value prediction model selection unit 140 determines whether there are any remaining control value prediction models that have not been processed in steps S401 to S406 (step S407). If there is an unprocessed control value prediction model (YES in step S407), return to step S401 and execute the processes of steps S401 to S406 on the unprocessed control value prediction model (for example, control value prediction model M2). do. If there is no unprocessed control value prediction model (NO in step S407), the process advances to step S408.

In step S408, the control value prediction model selection unit 140 selects each control value prediction model (control value prediction model Compare the average absolute errors calculated from M1, M2).

Then, the control value prediction model selection unit 140 selects, for each group, the control value prediction model with the smallest average absolute error in the comparison in step S408 as the control value model with high accuracy of the predicted control value (step S409), The group-by-group control prediction model selection process ends. In this example, it is assumed that in step S409, the control value prediction model M2 is selected for groups 1 and 3, and the control value prediction model M1 is selected for groups 2 and 4.

In the second embodiment, the control value prediction model calculation unit 150 calculates a plurality of control value prediction models (for example, control value prediction models M1 and M2) learned by the control value prediction model learning unit 120, as in the first embodiment. Measured data (current time information) of the plant 210 is inputted to calculate each predictive control value.

Furthermore, as a process specific to the second embodiment, the control value prediction model calculation unit 150 identifies the group to which the current measurement data belongs from the above measurement data. The performance data classification unit 310 may perform group identification of the current measurement data.

Then, the control value prediction model calculation unit 150 transmits the predicted control values of the control value model selected for each group in step S409 of FIG. 8 to the monitoring control system of the plant 210, according to the specified current group. . Applying the above-mentioned specific example, if the specified current groups are 1 or 3, the predicted control value of the control value prediction model M2 is transmitted, and if the specified current group is 2 or 4, the predicted control value of the control value prediction model M1 is transmitted. Send the predicted control value. The supervisory control system of the plant 210 then controls the operation of the flocculant injection pump of the plant 210 using the received predicted control value. Note that the supervisory control system of the plant 210 may not operate the received predicted control values as they are, but may operate them after confirmation by an operator.

As explained above, according to the operation management support device 300 according to the second embodiment, performance measurement data is classified into a plurality of groups, and depending on which group the current performance measurement data belongs to, A predictive control model to select and a predictive control value to be applied in the plant 210 can be determined. Specifically, for example, by setting Group 2 or 4 as a group for rain, group 1 as a group for winter, group 3 as a group for summer, etc., appropriate control values can be set according to annual water quality fluctuations. It becomes possible to select a prediction model. As a result, it can be expected that the amount of chemicals used will be further optimized than in the first embodiment. Further, since the operation management support device 300 according to the second embodiment also has the configuration of the operation management support device 100 according to the first embodiment, it is possible to obtain the same effects as the first embodiment. can.

Note that the present invention is not limited to the embodiments described above, and includes various modifications. For example, the embodiments described above are for explaining the present invention in an easy-to-understand manner, and are not necessarily limited to all the configurations. Further, the functions of each component may be integrated in any combination, for example, the control value prediction model learning section 120 and the control value prediction model selection section 140 may be combined. Further, the operation management support devices 100 and 300 are stored as software in a computer having a processor (CPU), a storage means such as a memory and an HDD, and a network interface (not shown), but these are not necessarily installed in the plant 210. It does not need to be located at the plant 210, and may be located at a location remote from the plant 210.

Furthermore, in the first and second embodiments, an example was shown in which the present invention was applied to chemical injection control in a water treatment plant, but the present invention also applies to blower control in a sewage treatment plant, temperature control in an electric power plant, and a garbage treatment plant. It can be widely applied to maintenance work that uses a predetermined type of control value in facilities that require explainability, such as, and the application is not particularly limited.

100,300 Operation management support device 110 Results DB
120 Control value prediction model learning section 130 Result prediction model learning section 140 Control value prediction model selection section 150 Control value prediction model calculation section 160 Result prediction model calculation section 170 Accuracy evaluation processing section 180 Display processing section 190 Processing result screen 210 Plant 220 Display Device 310 Actual data classification unit 320 Classification graph

Claims (12)

One or more control value prediction models that calculate a predicted control value that is a predicted value of a control value in the control by inputting measurement data indicating the performance of the plant regarding predetermined control regarding the operation of the plant;
a result prediction model that calculates a predictive control result that is a predicted value of the control result of the control using the predictive control value as input;
a control value prediction model calculation unit that inputs the measurement data and calculates a predicted control value for each control value prediction model from one or more of the control value prediction models;
a result prediction model calculation unit that inputs the predicted control value for each control value prediction model calculated by the control value prediction model calculation unit and calculates a predictive control result for each control value prediction model from the result prediction model;
Accuracy of the control value prediction model that calculates each predictive control result by comparing the predictive control result for each control value prediction model calculated by the result prediction model calculation unit with a predetermined control target value for the plant. an accuracy evaluation processing unit that determines a suitable control value prediction model to be applied to the plant by evaluating the
an output processing unit that outputs processing result information including at least the predicted control value calculated by the control value prediction model calculation unit and the result of the evaluation by the accuracy evaluation processing unit;
An operation management support device comprising: The operation management support device according to claim 1, wherein the control value prediction model is at least one of an empirical formula, a statistical model, or a physical model that reflects the knowledge of an operator. a control value prediction model learning unit that learns the control value prediction model using the measurement data;
A control value prediction model that selects a first control value prediction model with high accuracy of the predicted control value to be calculated from among the two or more control value prediction models based on the learning result by the control value prediction model learning unit. further comprising a selection section;
The operation management support device according to claim 1, wherein the control value prediction model selection unit applies the selected first control value prediction model to the operation of the plant. The accuracy evaluation processing unit is configured to evaluate the accuracy of the first control value more than the difference between the predictive control result calculated from the result prediction model using the predictive control value of the first control value prediction model and the control target value of the plant. When the difference between the predicted control result calculated from the result prediction model using the predicted control value of the second control value prediction model other than the control value prediction model and the control target value of the plant is smaller, the first The second control value prediction model is applied to the plant and determined to be a more suitable control value prediction model than the control value prediction model of , and the output processing unit is instructed to output the result of the determination. The operation management support device according to claim 3. If the difference between the predictive control result for each control value prediction model calculated by the control value prediction model calculation unit and the control target value of the plant is larger than a predetermined first threshold, the accuracy evaluation processing unit 4. The operation according to claim 3, further comprising: instructing the output processing unit to output a recommendation to re-learn a control value prediction model that has calculated the control prediction value used to calculate the predictive control result. Management support equipment. The accuracy evaluation processing unit evaluates the accuracy of the result prediction model based on the predictive control result calculated from the result prediction model using the predictive control value of the first control value prediction model and the measurement data of the plant. The operation management support device according to claim 3, wherein the operation management support device evaluates. further comprising a result prediction model learning unit that learns the result prediction model using the measurement data,
The accuracy evaluation processing unit determines that the difference between the predictive control result calculated from the result prediction model using the predictive control value of the first control value predictive model and the measured data of the plant is a predetermined second threshold value. 7. The operation management support device according to claim 6, wherein if the result prediction model is larger than , the output processing unit is instructed to output a recommendation to re-learn the result prediction model. The processing result information output by the output processing unit includes a trend graph of the measurement data, a predicted control value of the first control value prediction model, and a predicted control result of the result prediction model. The operation management support device according to claim 3. further comprising a performance data classification unit that classifies data into groups based on the degree of similarity between the measurement data,
The performance data classification unit classifies the predicted control values for each control value prediction model calculated by the control value prediction model calculation unit from the two or more control value prediction models into the groups,
The control value prediction model selection unit selects, for each group, a control value prediction model that has calculated the most accurate predicted control value among the predicted control values classified into the group by the performance data classification unit. selected as the first control value prediction model in
The control value prediction model calculation unit specifies the group to which current measurement data of the plant belongs, and applies the first control value prediction model corresponding to the specified group to the operation of the plant. The operation management support device according to claim 3. The operation management support device according to claim 7, wherein the result prediction model learning unit learns the result prediction model by machine learning using the measurement data. the plant is a water treatment plant;
The control value prediction model is a prediction model regarding injection of a predetermined drug into the water purification plant,
The result prediction model is a prediction model regarding the quality of treated water at the water purification plant,
The operation management support device according to claim 1, wherein the measurement data includes at least any one of data on water quality, water amount, and operation amount in the water purification plant. An operation management support method using an operation management support device that supports operator decision-making regarding predetermined control regarding plant operation, the method comprising:
The operation management support device includes:
one or more control value prediction models that calculate predicted control values that are predicted values of control values in the control by inputting measurement data indicating the performance of the plant;
a result prediction model that calculates a predictive control result that is a predicted value of the control result of the control using the predictive control value as input;
has
a control value prediction model calculation step in which the operation management support device inputs the measurement data and calculates a predicted control value for each control value prediction model from one or more of the control value prediction models;
A result in which the operation management support device inputs the predictive control value for each control value prediction model calculated in the control value prediction model calculation step and calculates a predictive control result for each control value prediction model from the result prediction model. a predictive model calculation step;
The operation management support device compares the predictive control results for each control value prediction model calculated in the result prediction model calculation step with a predetermined control target value for the plant, and calculates each predictive control result. an accuracy evaluation step of determining a suitable control value prediction model to be applied to the plant by evaluating the accuracy of the control value prediction model;
an output step in which the operation management support device outputs processing result information including at least the predicted control value calculated in the control value prediction model calculation step and the result of the evaluation in the accuracy evaluation step;
An operation management support method comprising:

Images